A Multi-Agent System for AI Coaching

In the first piece in this series I introduced two ideas that define what any agent can do: its environment, the world it perceives and operates in, and its action space, the set of actions available to it within that world. The scope and complexity of the environment and action space determine what type of AI is needed for a given use case. 

A chatbot answers what you put in front of it, and answers it in a similar way next month. A single-task agent does one narrowly defined job well. Ask an AI to draft an email, where the environment is a blank text box and the action space is a few hundred words, and a chatbot or single-task agent are often as good as anything more elaborate. 

But they are not built to coach. 

Coaching an enterprise workforce requires a multi-agent system because the problems are far more complex, interpersonal, and context-dependent and resolve over weeks or months rather than in a single coaching conversation. 

Consider Maya Chen, a newly promoted senior manager at a company she has never worked at before.¹ In her first hundred days she has to learn how the organization communicates and decides, earn the trust of seven direct reports she didn't hire, build credibility with peers whose priorities don't yet line up with hers, and deliver something visible while every relationship around her is still forming. Her environment is an entire company, and her action space runs from role-playing a hard conversation to setting goals to working through a problem in real time. Memory is the job because the quality of the guidance she gets depends on how well her coach understands the situation as it changes (i.e., what was said in yesterday's one-on-one, what is on her calendar tomorrow, who said what about whom, what she has already tried and how it landed).

As Maya’s example demonstrates, building an enterprise talent platform powered by an AI coach is a fundamentally different design problem than building a model that can answer any question on the internet. To develop talent at every part of an organization, an AI coach cannot be a feature bolted onto a general-purpose assistant, an AI layer added on top of a human coaching marketplace, or a document you upload with a coaching framework; it has to be an agentic system, purpose-built from the ground up.

¹Maya is a composite of a first-time manager, not a real user, and is an example of a population Valence has explicitly designed Nadia to serve.

The Agents Behind Nadia

We’ve built Nadia to be a multi-agent system. Nadia has the memory and context to coach Maya, but more powerfully, as a multi-agent system, Nadia can also serve as an AI-native platform for talent, performance, and learning for an entire organization.

To understand Nadia as a multi-agent system, let’s explore what happens when Maya opens Nadia and types: "Hi Nadia, I need to prepare for a difficult one-on-one conversation."

Instead of one model writing a reply, a set of specialized agents run in parallel, each handling a different part of the work.

Model orchestration comes first, and the user never sees it. Foundation models are the base layer, the raw reasoning and language that make everything else possible, but a system built on a single model is bounded by that model's general-purpose design, so Nadia directs work to the best model for any given task: one for chat, another for reasoning, another for memory operations, another for safety, and so on.

Situation analysis agents ask what Maya actually needs before trying to answer. Is this a request for a tactical script, a sounding board to think out loud, a short prep memo, or the visible edge of a pattern she keeps falling into with a particular member of her team? A general-purpose AI or single agent will quickly default to an answer, when the right coaching move is often creating productive friction by asking a better question back to clarify the underlying needs and goals.

Memory and context build Nadia's working model of Maya's world and run across three dimensions:

• Relational memory: the people around her and how those relationships actually function, including this report, her manager, and the peer she clashed with last month.
• Longitudinal memory: her goals, her progress, and the arc of her development across many conversations.
• Organizational context: the values, leadership frameworks, and talent processes specific to her company.

As a quick technical note, the base models are rapidly improving in their memory capabilities. Opus 4.8, the most capable model available as I write this, carries context across sessions so it can hold a multi-day project in mind from beginning to end; that continuity is real, it is useful, and it raises the floor for everything built on top. However, these models are designed to be general purpose, and the model memory tends to be summarized and abstracted. 

Coaching often depends on remembering exact language, and so memory has to answer a harder question: not what carried over from last time, but what exactly did this person say last time and how did they say it. It’s a memory built to know the person receiving coaching, their relationships, their role in the bigger picture of the organization.

For Maya's message, this is where Nadia recognizes and applies relevant context in ways general purpose models can’t. Nadia identifies which report is involved, the goal set two weeks ago in their 1:1, and the relevant part of the performance framework.

Coaching and personalization guide how Nadia decides to help. A personalization step shapes the response for Maya specifically, given her role, her tenure, and how she does her best thinking. It draws on the personality assessment Maya completed and, where the report has shared one, their collaboration profile in Nadia, so the guidance reflects not just how Maya works but how the person on the other side of the conversation does too. A coaching step selects from a library of validated approaches, such as goal-setting, role-play, and reflective questioning, and picks the one most likely to move her forward on this problem, in this moment.

Proactivity and tools turn the response into follow-through. Maya sees one answer, but behind it the system decides whether this should also produce something durable: a prep document to take into the meeting, a calendar hold, a reminder beforehand, a check-in afterward, or an update to the goal she set last session. 

Trust and safety agents run across all of the above rather than at the end of it. We use defense in depth: model-level safeguards, input checks for sensitive content and PII, system prompt defenses against biases that the underlying models may introduce, reasoning-stage evaluation of whether the planned response is appropriate, and output-stage review, each tuned to a different operating point. Many of the highest-value coaching moments touch difficult subjects, and we spend the most time on the hardest cases: a potential pre-PIP conversation, a legal or medical question, and anything touching harm or crisis. 

All of these agents coordinate the response in seconds, and Maya sees one response, in one voice, that already knows what she is working on, what she has tried, and what is likely to help right now; in this case, an offer to role-play the conversation against the goals she set last time.

This is what happens in the conversation, but in practice, more agents run between Maya's conversations than during them. A set of agents examines what was said in the context of everything that came before, distilling what is new, what changed, and what a coach needs to remember next time, then updates Maya’s plan, profile, and what the system knows about which coaching moves land for her.

Nadia’s coaching in month twelve is even better than in week one

The architecture matters because of what it makes possible over time. In week one, Nadia knows the basics: Maya's role, her company's leadership framework, the goals she set in onboarding. By month three, Nadia knows her people, with working profiles of each direct report: who needs context before recommendations, who responds to questions rather than answers, who is quietly disengaging. The system remembers which team member Maya is worried about and why. By month twelve, Nadia knows Maya's year: the patterns, the commitments, the relationships, the moments where things got hard and where she grew through them. The picture deepens because the environment grew richer, the action space grew deeper, and the system learned from everything that came before.

Now, we’ve looked here at just a single coaching relationship between Nadia and Maya. But the most exciting part is that Nadia doesn’t just coach one person, Nadia can coach the entire organization at the same time. For HR leaders, it’s an AI platform for talent, performance, and learning that translates strategy from executive decks into day-to-day behaviors and decisions across an entire company. 

‍

Jeff Dalton is Head of AI and Chief Scientist at Valence and a professor at the University of Edinburgh, where his research group focuses on conversational AI, agentic information access, and neural ranking. He is a Turing AI Fellow and the author of more than 100 research papers. Prior to Valence, he led development of language understanding capabilities for Google Assistant and built next-generation knowledge graphs for Google Search.

‍